How to produce ast edits or patches?

[nashid]

We were able to run the training and the inference (i.e. evaluation) script. However, we can only find the accuracy metrics. Is it also possible also get the AST edit operations in the tree?

./eval.sh
use cpu
loading HOPPITY from /Users/zhutao/lab/hoppity
Namespace(act_func=‘tanh’, ast_fmt=‘shift_node’, att_type=‘inner_prod’, batch_size=10, beam_agg=False, beam_size=3, comp_method=‘mlp’, data_in_mem=False, data_name=‘contextmltttttzzz’, data_root=‘/Users/zhutao/lab/data/small_astPKL’, dataset_stats=False, dropbox=None, dropout=0, end_epoch=10000, eval_dump_folder=‘~/eval_dump/’, gnn_msg_dim=128, gnn_out=‘last’, gnn_type=‘s2v_multi’, grad_clip=5, hinge_loss_type=‘sum’, init_model_dump=None, iters_per_val=100, lang_dict=‘None’, latent_dim=128, learning_rate=0.001, loc_acc=True, loc_given=False, loss_file=‘loss.txt’, max_ast_nodes=500, max_lv=4, max_modify_steps=1, max_token_len=100, min_lr=1e-06, mlp_hidden=256, msg_agg_type=‘sum’, neg_samples=1, num_cores=4, num_epochs=10000, op_acc=False, op_breakdown=False, op_given=False, output_all=True, penalize_unknown=False, phase=None, rand=False, raw_srcs=None, readout_agg_type=‘sum’, resampling=True, rnn_cell=‘gru’, rnn_layers=2, sample_list=None, save_dir=‘/Users/zhutao/lab/data/small_trainingResult’, seed=19260817, sibling_acc=False, sqr_data=None, start_epoch=0, target_model=‘/Users/zhutao/lab/data/small_trainingResult/epoch-0.ckpt’, test_pct=0.1, topk=3, train_pct=0.8, type_acc=False, val_acc=True, val_pct=0.1, vocab_type=‘fixes’)
====== begin of s2v configuration ======
| msg_average = 0
======   end of s2v configuration ======
not loading cuda jagged ops
not loading cuda metric ops
loading cooked asts and edits
51it [00:00, 3214.80it/s]
51 samples loaded.
loading vocab from /Users/zhutao/lab/data/small_astPKL/type_vocab.pkl
256 types of nodes in total.
train set has 41 samples
val set has 5 samples
test set has 5 samples
loading /Users/zhutao/lab/data/small_trainingResult/epoch-0.ckpt
Beam agg False
0it [00:00, ?it/s]----- replace_val
1it [00:03,  3.33s/it]
total accuracy 0.0000
location accuracy 0.0000
value accuracy 0.2000
number of unique edits 0 total samples 5
number of unknowns 0

[elizabethdinella]

Hi, yes this is possible. The following lines in eval.py contain a list of the edits predicted as GraphEditCmd objects.

https://github.com/AI-nstein/hoppity/blob/master/gtrans/eval/eval.py#L145-L149

[elizabethdinella]

Hi @nashid. Any update? Did this work out for you?

[Legitao]

Hi @elizabethdinella , thanks for your help, we are able to get the edits now:

Is there a way to find the path(e.g. /items/0/expression/arguments/1/type) to the bug location in JSON according to node_index here?